ddecosmo/lanternfly_training_dataset

Dataset Card for {{ pretty_name | default("Dataset Name", true) }}

This dataset is the training dataset for 24-679 Project 1: Lanternfly Tracker It is composed of 360 original lanternfly photos, 150 original photos with no lanternflies, and 800 original photos from nature, urban, and other insect datasets listed below.

These were augmented 50X to 65.1k augmented images.

Dataset Details

Dataset Sources [optional]

Original Lanternfly Datasets rlogh/lanternfly-data: Original Lanternfly Dataset, 229 unmarked photos rlogh/lanternfly_swatter_training: Dataset with geolocal data: 165 photos

Original Negative Datasets: rlogh/negativesirl: Negatives dataset, images of outdoor environements and people with no lanternflies. 107 photos

Total: 501 original images

Imported Datasets uoft-cs/cifar100: General image classifier, no insect class AI-Lab-Makerere/beans: Foliage with no insects Francesco/insects-mytwu: Insect Images

Total: 800 additional images imported

Uses

These images were used to train the EfficientNetB1 model, ddecosmo/lanternfly_classifier, on how to classify images as containing or not containing lanternflies.

Direct Use

The direct use is identifying photographs containing lanterflies so this could be used for tracking purposes.

Out-of-Scope Use

In future, this model could be adapted to identify other types of insect within this dataset.

Dataset Structure

This dataset consists of two splits An original split with 1.3k photos An artificial split with 65.1k photos

The tasks fall into 3 categories based on the building pictured

1. Lanternflies, all original photos
2. Other Insect, all 3rd party datasets
3. No insect, original photos and 3rd party datasets

Dataset Creation

Source Data

This data is sourced by the creators, Devin and Rumi for all original photos

Additional datasets can be found here, uoft-cs/cifar100 AI-Lab-Makerere/beans Francesco/insects-mytwu

Data Collection and Processing

Original datasets were collected using the mobile phones of the authors.

Additional datasets were recommended by Gemini AI and then validated as fitting the purpose, type, and scope of this process. uoft-cs/cifar100: This is a general image identifier with no insect class. Used for no insect for generalizability AI-Lab-Makerere/beans: This dataset is focused on vegetation with and without disease, this is used to train the model to recognize vegetation without insects/lanterflies. Francesco/insects-mytwu: This is an object detection dataset used for identifying insects as subjects, not including lanterflies. We are using it train a seperate non-lanternfly insect class.

Who are the source data producers?

Original data was produced by the authors.

Additional datasets were produced by, uoft-cs/cifar100: Created by University of Toronto Computer Science AI-Lab-Makerere/beans: Created by AI Lab Makere Francesco/insects-mytwu: Created by Fanscesco Sovrano

Bias, Risks, and Limitations

The main risk of this dataset is the lanternfly split. It contains only images of singular lanternflies on the ground. Normally on concrete or asphalt. This severly limits the scope of the environments these creatures appear in. Incorporating blob detection or YOLO into future models could mitigate this by focusing on the subject.

Recommendations

This is a large dataset, and has been shown to accurately classify lanternflies, but there are many edge cases when it does not work correctly. In order to take this into account, using new types of models with subject detection can make use of the many images while improving model accuracy.